Machine Learning
Research on Machine Learning: explanations, evidence, limitations and connections between papers.
88 results
Specify the event and clock before trusting a decision
A game planner can misprice one reward, a voice assistant can finish an outdated task, and a video monitor can alert at the wrong moment. An earnings-call benchmark can also start measuring a return before the input it describes is ready. Each case depends on where an event boundary is placed. The studies make those boundaries visible through transition substitution, task-version checks, alert rules, and declared price clocks. Their findings answer different questions under different protocols; the useful first step is to identify the event, the relevant state, and the clock before interpreting a result.
Changing model settings can reduce prediction coverage
An audit finds that changing inference settings can reduce conformal coverage on finite-label tasks, though effects vary by model and dataset.
When whole-task audits misjudge component checkers
A theory and ECG study show how whole-task audits can weaken component checkers, while retaining their failures preserved performance in tested settings.
A gate checks whether model updates really improve
BRIDGE checks model updates on separate data and reduces false improvements in tested tabular tasks; its population bound is synthetic.
When compilers can safely move MXFP4 scales
A compiler method proves when shared MXFP4 scales can move across a binary32 reduction without changing its result.
Agents can improve research code, but many tasks stump all
FrontierCS 2 tests agents improving research code; its top agent passed 58% of runs across 48 tasks under an all-metrics rule.
A satellite system brings stronger rainfall detection into real time
Huayu-Global combines three satellite streams and detects precipitation better overall than seven products in matched station comparisons.
For LLM inference, lower precision beats 2:4 above the floor
A comparison of 2:4 sparsity and quantization finds dense formats more accurate above the lowest supported precision in two 8B models.
A bit-based method compares context with model weights
An information-based method compares context with trained weights; issued certificates matched measured winners in a designed transformer test.
A flow-wide method improves low-temperature generation
Bayes-Consistent Tempering aligns training and sampling for flow language models, improving low-temperature results in tested math and Sudoku tasks.
A memory branch improves GraphCast’s longer-range forecasts
A memory correction branch lowered GraphCast’s weighted forecast loss across ten-day forecasts from 122 starts in 2023.
Match parallel generation to the work that slows it down
A phone agent can wait at every screen action, while a text generator can spend time proposing, checking, or reconstructing its output. Parallel work helps when it reaches the costly part of that particular workflow. Across these studies, moving or overlapping work produces reported throughput or latency gains in specific workloads and serving setups. JetSpec++ also reports a serving gain, while its predictions of future context remain estimates until verification. The results support judging each method by its local outcome, without forming a shared speed ranking.
Studies test final self-improving systems more than the loop behind them
A study of papers finds repeated tests of full improvement loops less common than evaluations of final systems.
A new method tracks how training targets change predictions
The target-response Jacobian closely tracks refitted prediction shifts after training-target changes on three tested tabular datasets.
A score can rank when CT systems fail without warning
A prospective score ranked silent failure in two learned CT systems, with evidence limited to simulated access tests.
Racer checks whether GPU kernel rewrites preserve their function
Racer symbolically checks GPU kernel rewrites and exposes bugs missed by sampled tests, within the coverage of its production-kernel study.
Shared anchors let language-model monitors transfer across models
Shared-anchor probes reuse behavior monitors across language models, but transfer varies by task and target.
Two-class pretraining still supports multiclass table predictions
A TabICLv2 study finds two-class pretraining retains multiclass prediction, with results bounded by training recipe differences.
Same-size students struggle to match model harnesses
Trace-trained students lag behind harness-assisted models on tested puzzles, while greater capacity or simpler tasks can narrow the gap.
Synthetic training can fail where generated data leave gaps
A theory of synthetic training shows when coverage supports real-data error guarantees and how uncovered regions can defeat them.
A speaker’s ordinary voice can trigger a speech-model backdoor
A two-phase method ties a target voice to chosen outputs, succeeding selectively in two tested speech models.
HarnessPro improves agent success at lower cost
HarnessPro evolves AI agent software with execution feedback and generated tests; one terminal-task comparison improved success at lower cost.
Small adapters improve brain-response predictions across tested models
BrainLoRA trains small adapters to improve brain-response predictions across tested models, with transfer evidence limited to narrative films.
Text-to-image models can learn without text-only pretraining
Halcyon learns from paired captions and images; tests found no benefit from text-only pretraining in the specific settings studied.